JP2707564B2 - Audio coding method - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech coding method, and
The present invention relates to a method for improving the quality of encoded audio when compressing to around bps. [Prior Art] In order to transmit an audio signal through a broadband cable, the audio signal is sampled, quantized, and converted into a binary digital code, thereby performing PCM transmission. On the other hand, when constructing a communication network using a dedicated digital line, reduction of communication cost is a very important issue, and the amount of information of a voice signal reaching 60 Kbps is too large, so that it cannot be transmitted as it is. Therefore, information compression (that is, low bit rate encoding) of the audio signal for transmission has become necessary. As an audio encoding method for compressing an audio signal at around 8 Kbps, a method is known in which audio is separated into spectral envelope information and sound source information, and each is encoded. inside that,
The so-called PARCOR (Partial Autocorrelation: Partial Autocorrelation) method, in which the sound source information is modeled by a single pulse train and white noise, can be encoded at a low bit rate, but on the other hand, the quality deteriorates. large. On the other hand, the multipulse method (for example, Ozawa, et al., "Quality Improvement of Multipulse Driven Speech Coding", Speech Research Meeting of the Acoustical Society of Japan)
S83-78 (1984.1), or residual compression method (see Asakawa, et al., "Speech synthesis method using residual information," Proceedings of the Acoustical Society of Japan, 3-1-7 (Showa 59.10)) There is. As a residual compression method, for example, a method described in JP-A-61-296398 has been proposed.
No. 1419 and Japanese Patent Application No. 61-35148. In these methods, as the expression of the sound source is refined,
The quality is improved compared to the PARCOR method. [Problems to be Solved by the Invention] In the above-described conventional technique, a plurality of pulse trains as sound sources are generated independently for each frame on a fixed basis.
Here, the frame is a time unit for analyzing the voice, and is usually set to about 20 ms. By the way, it is assumed that the audio waveform has been sampled and converted into a series of sample values x _i . Let x _t be the present and then p sample values going back to the past are ｛x
_ti ｝, (i = 1,..., p). Here, it is assumed that the speech waveform can be approximately predicted from the past p samples. Since the simplest prediction is linear prediction, it is considered that current values are approximated by multiplying each of past sample values by a certain coefficient and adding them. In this case, the difference between the realized value x _t at the current point t and the predicted value y _t, and the prediction error epsilon. This prediction error ε is called a prediction residual or simply a residual. The predicted residual waveform of the speech waveform is considered to be the sum of two types of waveforms. One of them is a so-called error component, the amplitude of which is not so large, and is close to a random noise waveform. The other one is an error when a pulse due to vocal cord vibration is applied to the input, and the prediction is largely out of order, resulting in a residual waveform having a large amplitude. This residual component is
It appears periodically and repeatedly according to the periodicity of the sound source. Speech is roughly classified into a section having periodicity (voiced sound) and a section having no remarkable periodicity (unvoiced sound), and accordingly, the prediction residual waveform also has periodicity in the voiced sound portion. doing. On the other hand, the pulse train generated in the multi-pulse method or the residual compression method can be regarded as an approximation of the residual,
The voiced part should have periodicity. However, since these pulse trains are generated independently of the preceding and succeeding frames, the relative positional relationship of the pulse trains is shifted for each frame, and the periodicity may be disturbed. When speech is synthesized using such a pulse train as a sound source,
There was a problem that the sound quality was degraded, called "gorogoro." An object of the present invention is to improve such a conventional problem and prevent deterioration of sound quality due to disturbance in periodicity of each frame with respect to a pulse train generated by a multi-pulse method or a residual compression method. It is to provide a speech coding system. [Means for Solving the Problems] In order to achieve the above object, the speech coding method of the present invention is to determine whether a voiced frame is switched from an unvoiced frame, immediately after a voiced frame is continuous, or is a voiceless frame. A first sound source pulse generating means for generating a sound source pulse immediately after switching from the unvoiced frame to a voiced frame, and a second sound source pulse generating means for generating a sound source pulse when the voiced frame continues. It is characterized by comprising: a second sound source pulse generating means; and a third sound source pulse generating means for generating a sound source pulse in the case of the unvoiced frame. [Operation] In the present invention, the position of the pulse train of the next frame is estimated based on the pitch period based on the first generated pulse train, and a new pulse train is generated in the vicinity of the estimated position. Hold. That is, the voice cycle of the voiced sound corresponds to the pitch cycle that is the reciprocal of the pitch frequency that is the pitch of the voice. Since the change in pitch is relatively gentle, it can be regarded as substantially constant within one frame. Therefore, in the first reference frame, for example, in the first frame in which voice is switched from unvoiced sound to voiced sound, after generating a sound source pulse train based on a certain standard according to the related art, the next frame is sequentially generated based on the generated sound source pulse train. The method of estimating the position of the sound source pulse train in and generating the sound source pulse train is used. In the multi-pulse method and the residual compression method, since the number of sound source pulses is small, generally generated sound source pulse trains are grouped together for each pitch period. Therefore, the position advanced in the time axis direction by the pitch period with respect to the sound source pulse train in the last pitch period of the frame is set as the position of the leading pulse train of the next frame. In this way, the periodicity of the pulse train between two frames is maintained. In the next frame, based on this position,
A first sound source pulse train is generated near this position. As a result, the disturbance of the periodicity between the frames is eliminated, the deterioration of the sound quality can be prevented, and the optimum sound source pulse train based on the pulse train generation standard can be obtained. EXAMPLES Hereinafter, examples of the present invention will be described in detail with reference to the drawings. FIG. 1 is a block diagram showing a case where the speech coding method of the present invention is applied to a speech coding device (speech CODEC) using a residual compression method, wherein (a) is a coding unit;
(B) is a decoding unit. As shown in FIG. 1 (a), the encoding unit of the present invention includes a buffer memory 1 for storing a digital audio signal, a linear prediction circuit 3 for performing linear prediction, and an inverse filter controlled using parameters 4. 5, a pitch extraction circuit 7 for extracting a pitch using a residual correlation method or the like, and a voiced / unvoiced determination circuit 9
A sound source generating unit 11 for generating a sound source pulse in accordance with a voiced / unvoiced determination result; and a quantization encoding circuit 13.
Further, as shown in FIG. 1 (b), the decoding unit of the present invention comprises a decoding circuit 16 for separating an input signal into four types of parameters.
And a buffer memory 19 for storing the decoded spectral parameters, a pitch pulse, a voiced / unvoiced determination result, and a sound source information as input, a sound source pulse reproducing circuit 17 for generating a sound source pulse, and a delay in the sound source pulse reproducing circuit 17. And a synthesizing filter 20 that uses the result as a coefficient. In FIG. 1 (a), at the time of encoding, a digitized audio signal is stored in a buffer memory 1 for one frame, and a parameter (for example, a partial (Autocorrelation coefficient) 4. Next, an inverse filter 5 is configured using the parameter 4 as a coefficient, and the audio signal 2 is input to the inverse filter 5 to obtain a residual signal 6. Pitch extraction circuit 7
Is based on the residual correlation method or AMDF (Average Magnitude Differenti
al Function) method and other well-known methods,
A pitch period 8 of a frame is extracted using the residual signal 6 as an input. The voiced / unvoiced determination circuit 9 outputs a determination result 10a indicating whether the frame is a voiced frame or an unvoiced frame, and a signal 10b indicating that the frame has been switched from an unvoiced frame to a voiced frame. The sound source generation unit 11 is newly provided according to the present invention, and includes a voice unvoiced determination result.
A sound source pulse is generated according to 10a and the switching signal 10b, and the information 12 is output. The quantization encoding circuit 13 receives the spectrum parameter 4, the pitch period 8, the voiced / unvoiced determination result 10a, and the sound source information 12, quantizes it to a predetermined number of bits, and converts the result 14 converted into a predetermined format into a digital line. Send to 15. In FIG. 1 (b), at the time of decoding, digital data 14 received from a digital line 15 is decoded.
, The four parameters shown in (a) (pitch period 8 ', sound source information 12', voiced / unvoiced judgment result 10a ',
It is separated into spectral parameters 4 '). Three of the above four parameters (decoded pitch period 8 ', voiced / unvoiced determination result 10a', sound source information 12 ')
A target sound source pulse 18 is obtained by a sound source pulse reproducing circuit 17 having the input as an input. On the other hand, only one of the four parameters (decoded spectrum parameter 4 ') is stored in the buffer memory 19, and after correcting the delay in the excitation pulse reproducing circuit 17, the buffer memory 19 Is used as a coefficient of the synthesis filter 20. By inputting the sound source pulse 18 to the synthesis filter 20, a synthesized speech 21 can be obtained as the output. FIG. 2 is a functional block diagram of a sound source generation unit in FIG. As shown in FIG. 2, the sound source generation unit 11 includes a switching control unit 31 for switching control when voice is switched from unvoiced to voiced, and a buffer memory 111 for storing a residual signal.
When the voice is switched from unvoiced to voiced, the pulse extraction position determination unit 112 for determining the pulse extraction position, and the head address of the representative residual determined in the previous frame is converted to the address of the buffer memory 111 and stored. And a pulse extraction position determination unit 32 for determining a pulse extraction position when voiced frames are continuous.
And a sound source extraction unit 115 for extracting a sound source from the head address and buffer memory 111, and an unvoiced sound source generation unit 116 for generating an unvoiced sound source. Since the speech coding method according to the present embodiment is related to the generation of a voiced frame sound source, the voiced / unvoiced determination result 10a indicates voiced, and the pitch period 8 is assumed to have a fixed value (hereinafter referred to as pitch The value of the cycle is NPTCH). First, when the voiced / unvoiced switching signal 10b indicates that it is immediately after switching from unvoiced to voiced, control is performed by a signal from the switching control unit 31 to control the pulse extraction position determining unit (I) 11
Move on to 2. The function of the sound source generation unit 11 when controlled here is the same as that of the conventional residual compression method (for example, the above-mentioned publication (Japanese Unexamined Patent Publication No.
No. 296398), which is the same as the residual compression method described as the second method. That is, consecutive LN residual pulses are extracted for a representative pitch section (here, LN is the number indicated by the value of the number of extracted pulses 113). Further, as described in the above-mentioned Japanese Patent Application No. 60-241419, when the decoding residual of the previous frame and the representative residual of the current frame are interpolated at the time of decoding, the representative pitch section is the current frame. To include the last point of That is, the pulse extraction position determination unit (I) 112 calculates the following equation. Here, i satisfies the following conditional expression. iFRM−NPTCH + 1 ≦ i ≦ iFRM (2) where x _j is the residual pulse amplitude at address j,
It is read from the buffer memory 111. Note that the buffer memory 111 is a ring buffer and stores the residual of the previous frame and the current frame. Also, iFRM is the frame length, and LN is the value of the number of extracted pulses 113. For example, in order to obtain the amplitude information and the position information of the next residual pulse to be interpolated, the pulse extraction position test unit 112 first obtains the cumulative amplitude value using the above equations (1) and (2). Now, the current frame length is stored in the buffer memory 111 as 0 to 159.
Is assigned, and if there are 20 consecutive residual pulses for the representative pitch section, the next representative pitch section is determined to include the last point of the current frame, and the above equation (2) The position i to be determined is determined within a section smaller than the frame length and larger than a section smaller by the pitch period than the frame length. Then, the head address is obtained from the accumulated amplitude value calculated by the above equation (1), and interpolation is performed by taking out 20 residual pulses from the address from the buffer memory 111. Assuming that i which gives the maximum value of AMP (i) calculated by the above equation (1) is i ₀ , i ₀ is the head address 114a of the representative residual. When the head address 114a is sent to the sound source extraction unit 115,
The LN residuals are read from the buffer memory 111 from the start address, and are transmitted to the subsequent stage as sound source information 12. Next, a case where the voiced / unvoiced switching signal 10b is not immediately after switching from unvoiced to voiced, that is, a case where voiced frames are indicated as being continuous will be described in detail. At this time, the control is transferred to the pulse extraction position determination unit (II) 32 by the signal from the switching control unit 31. The buffer memory 111 stores the residual of two frames. Addresses -iFRM + 1 to 0 correspond to the previous frame, and 1 to iFRM correspond to the current frame. Also, in the start position memory 30, the start address i ₀ of the representative residual determined in the previous frame is converted into an address on the buffer memory 111 (i ₀ ′ = i ₀ −IFRM) and stored.
The leading position of the representative residual of the current frame is determined as follows based on i ₀ ′. In the above equation (3), STADRS ₁ ,..., ST
ADRS _N corresponds to the head address for interpolating the representative residual at the time of decoding, and STADRS _N is the one in the last pitch section in the current frame, that is, the head address of the representative residual, and Become like i ₀ = STADRSN _N (4) In this way, the representative residual start address of the current frame can be obtained very easily from the representative residual start address of the previous frame. However, since the pitch cycle NPTCH is an average pitch cycle of the current frame, there is a possibility that the pitch cycle NPTCH has an error from the actual pitch position. Therefore, in order to determine the position more precisely, the following is performed. First, a short-term cross-correlation value is defined by equation (5). i ₀ ′ + NPTCH−D ≦ i ≦ i ₀ ′ + NPTCH + D (6) Here, D (> 0) is a value determined by pitch fluctuation and the like, and COR represents a cross-correlation value. In the above equation (6),
This shows that the existence range of the head address of the first excitation pulse train of the current frame is in the range obtained by adding the head address of the representative residual of the previous frame in consideration of the fluctuation of the pitch period. Number of pulses extracted from start address
This is to calculate the amplitude accumulated value of LN residual pulses, and the correlation value becomes the maximum value if the phases match. The first start address is obtained by the following equation. In the above equation (7), a position i where the correlation value becomes highest is detected near the position distant from the representative residual of the previous frame by NPTCH. Hereinafter, it is sufficient to replace i ₀ ′ with STADRS ₁ and obtain STADRS ₂ in the same procedure, and sequentially obtain up to STADRS _N (= i ₀ ). The above equation (1) can also be used to determine STADRS _n (where n is an arbitrary integer). That is, assuming that the range of i in the above equation (1) is the equation (6), the following equation (8) is derived. Hereinafter, the same procedure is used to obtain up to STADRS _N (= i ₀ ). The head address (i ₀ ) 114b of the representative residual determined by one of the methods described above is sent to the sound source extraction unit 115. At the time of decryption, the conventional method (for example, the aforementioned Japanese Patent Application No.
Thus, the excitation pulse is reproduced while interpolating the representative residual and the decoding residual of the previous frame. At this time, since the interpolation corresponding point address is the representative residual position itself of the previous frame, it is not necessary to transmit the address again. The sound source pulse generation unit 11 shown in the present embodiment can be easily realized by an adder, a correlator, a comparator, and the like, as described in detail above. The same function can be realized by a general-purpose microprocessor. In the current frame, when the voice unvoiced determination result 10a is unvoiced, the control is switched to the unvoiced sound source generating unit 116 by the control signal from the switching control unit 31. The operation of the unvoiced sound source generation unit 116 generates a sound source pulse irrespective of the pitch period, for example, as in a method proposed in the past (for example, see Japanese Patent Application No. 61-35148). FIG. 3 is a time chart for explaining the effect of the present invention. FIG. 3A is a waveform diagram showing an input sound source waveform 41, a residual waveform 42, a representative residual waveform 43a, and a composite waveform 44a according to a conventional method, and FIG. 3B is an input waveform according to the present embodiment. Audio waveform 41, residual waveform 42, representative residual 43b, and composite waveform 44b
FIG. The input voice waveform 41 has the same waveform in both (a) and (b), and the waveform 42 of the residual signal of the inverse filter 5 also has the same waveform. In the conventional method, since the representative residual (after decoding) is extracted independently for each frame, as shown in the waveform 43a,
In the frame # 3, the displacement of the representative residual occurs, and the periodicity is disturbed. Arrows indicate the shift width. As a result, as shown in FIG.
In the case of 44a, the amplitude is attenuated at the position where the displacement has occurred, resulting in deterioration of the sound quality. In the case of the present embodiment, as shown in FIG. 3 (b), when voiced frames are consecutive, the representative residual (after decoding) 43b extracted subordinately with reference to the representative residual position of the previous frame.
Becomes The representative residual 43b has no displacement, and therefore the composite waveform 44b has no attenuation, is natural, and has improved sound quality as compared with the conventional system shown in FIG. 3 (a). [Effects of the Invention] As described above, according to the present invention, when a voiced sound is continuous, a sound source pulse train is generated without disturbing the periodicity of the original sound, so that the periodicity is caused by disturbance. Deterioration of sound quality can be prevented, and the quality of encoded sound can be improved.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a block diagram of a speech encoding system showing one embodiment of the present invention, FIG. 2 is a block diagram of a sound source generation unit in FIG. 1, and FIG. 3 is an effect of the present invention. 7 is a waveform time chart for explaining the present invention. 1, 19, 111: buffer memory, 3: linear prediction circuit, 5: inverse filter, 7: pitch extraction circuit, 9: voiced / unvoiced discriminator, 11: sound source generator, 17: sound source pulse reproducer, 20: synthesis filter, 31 :
Switching control unit, 112, 32: pulse extraction position determination circuit, 30: head position memory, 116: unvoiced sound source generation unit, 115: sound source extraction unit, 6:
Residual signal, 12: sound source information, 21: synthesized voice, 43a, b: representative residual waveform, 44a, b: synthesized waveform, 42: residual waveform, 41: input voice waveform.

   ────────────────────────────────────────────────── ─── Continuation of front page    (72) Inventor Toshiro Suzuki               1-280 Higashi Koikebo, Kokubunji-shi, Tokyo                 Central Research Laboratory, Hitachi, Ltd.                (56) References JP-A-59-65897 (JP, A)                 JP-A-60-162300 (JP, A)                 JP-A-60-235200 (JP, A)                 JP-A-61-7899 (JP, A)                 JP-A-62-38500 (JP, A)

Claims (1)

(57) [Claims] The audio signal is analyzed for each frame, and is separated into spectral envelope information and sound source information, and it is determined whether the audio signal is voiced or unvoiced. In a voiced frame, a plurality of pulses are used as a sound source for one pitch period. Means for determining whether the voiced frame is switched from an unvoiced frame or whether voiced frames are continuous, and a sound source pulse immediately after switching from the unvoiced frame to a voiced frame. First sound source pulse generating means for generating a sound source pulse, second sound source pulse generating means for generating a sound source pulse when the voiced frame is continuous, and third sound source pulse generating for the unvoiced frame.
Sound source pulse generating means, wherein the second sound source pulse generating means determines the sound source pulse position of the current voiced frame by the pitch cycle based on the sound source pulse position of the voiced frame immediately before the current voiced frame. And generating an excitation pulse train near the determined position. 2. 2. The speech coding method according to claim 1, wherein a correlation method is used to determine a sound source pulse position of the current voiced frame.